1 /* Copyright (c) 2012. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 // 1./ check how and where a new VM is added to the list of the hosts
8 // 2./ Diff between SIMIX_Actions and SURF_Actions
9 // => SIMIX_actions : point synchro entre processus de niveau (theoretically speaking I do not have to create such SIMIX_ACTION
13 // MSG_TRACE can be revisited in order to use the host
14 // To implement a mixed model between workstation and vm_workstation,
15 // please give a look at surf_model_private_t model_private at SURF Level and to the share resource functions
16 // double (*share_resources) (double now);
17 // For the action into the vm workstation model, we should be able to leverage the usual one (and if needed, look at
18 // the workstation model.
20 #include "msg_private.h"
21 #include "xbt/sysdep.h"
24 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(msg_vm, msg,
25 "Cloud-oriented parts of the MSG API");
28 /* **** ******** GENERAL ********* **** */
30 /** \ingroup m_vm_management
31 * \brief Returns the value of a given vm property
34 * \param name a property name
35 * \return value of a property (or NULL if property not set)
38 const char *MSG_vm_get_property_value(msg_vm_t vm, const char *name)
40 return MSG_host_get_property_value(vm, name);
43 /** \ingroup m_vm_management
44 * \brief Returns a xbt_dict_t consisting of the list of properties assigned to this host
47 * \return a dict containing the properties
49 xbt_dict_t MSG_vm_get_properties(msg_vm_t vm)
51 xbt_assert((vm != NULL), "Invalid parameters (vm is NULL)");
53 return (simcall_host_get_properties(vm));
56 /** \ingroup m_host_management
57 * \brief Change the value of a given host property
60 * \param name a property name
61 * \param value what to change the property to
62 * \param free_ctn the freeing function to use to kill the value on need
64 void MSG_vm_set_property_value(msg_vm_t vm, const char *name, void *value, void_f_pvoid_t free_ctn)
66 xbt_dict_set(MSG_host_get_properties(vm), name, value, free_ctn);
69 /** \ingroup msg_vm_management
70 * \brief Finds a msg_vm_t using its name.
72 * This is a name directory service
73 * \param name the name of a vm.
74 * \return the corresponding vm
76 * Please note that a VM is a specific host. Hence, you should give a different name
80 msg_vm_t MSG_vm_get_by_name(const char *name)
82 return MSG_get_host_by_name(name);
85 /** \ingroup m_vm_management
87 * \brief Return the name of the #msg_host_t.
89 * This functions checks whether \a host is a valid pointer or not and return
92 const char *MSG_vm_get_name(msg_vm_t vm)
94 return MSG_host_get_name(vm);
98 /* **** Check state of a VM **** */
99 static inline int __MSG_vm_is_state(msg_vm_t vm, e_surf_vm_state_t state)
101 return simcall_vm_get_state(vm) == state;
104 /** @brief Returns whether the given VM has just reated, not running.
107 int MSG_vm_is_created(msg_vm_t vm)
109 return __MSG_vm_is_state(vm, SURF_VM_STATE_CREATED);
112 /** @brief Returns whether the given VM is currently running
115 int MSG_vm_is_running(msg_vm_t vm)
117 return __MSG_vm_is_state(vm, SURF_VM_STATE_RUNNING);
120 /** @brief Returns whether the given VM is currently migrating
123 int MSG_vm_is_migrating(msg_vm_t vm)
125 return __MSG_vm_is_state(vm, SURF_VM_STATE_MIGRATING);
128 /** @brief Returns whether the given VM is currently suspended, not running.
131 int MSG_vm_is_suspended(msg_vm_t vm)
133 return __MSG_vm_is_state(vm, SURF_VM_STATE_SUSPENDED);
136 /** @brief Returns whether the given VM is being saved (FIXME: live saving or not?).
139 int MSG_vm_is_saving(msg_vm_t vm)
141 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVING);
144 /** @brief Returns whether the given VM has been saved, not running.
147 int MSG_vm_is_saved(msg_vm_t vm)
149 return __MSG_vm_is_state(vm, SURF_VM_STATE_SAVED);
152 /** @brief Returns whether the given VM is being restored, not running.
155 int MSG_vm_is_restoring(msg_vm_t vm)
157 return __MSG_vm_is_state(vm, SURF_VM_STATE_RESTORING);
162 /* ------------------------------------------------------------------------- */
163 /* ------------------------------------------------------------------------- */
165 /* **** ******** MSG vm actions ********* **** */
167 /** @brief Create a new VM with specified parameters.
171 msg_vm_t MSG_vm_create(msg_host_t ind_pm, const char *name,
172 int ncpus, int ramsize, int net_cap, char *disk_path, int disksize)
174 msg_vm_t vm = MSG_vm_create_core(ind_pm, name);
177 s_ws_params_t params;
178 memset(¶ms, 0, sizeof(params));
179 params.ramsize = ramsize;
180 params.overcommit = 0;
181 simcall_host_set_params(vm, ¶ms);
184 /* TODO: Limit net capability, take into account disk considerations. */
190 /** @brief Create a new VM object. The VM is not yet started. The resource of the VM is allocated upon MSG_vm_start().
193 * A VM is treated as a host. The name of the VM must be unique among all hosts.
195 msg_vm_t MSG_vm_create_core(msg_host_t ind_pm, const char *name)
197 /* make sure the VM of the same name does not exit */
199 void *ind_host_tmp = xbt_lib_get_elm_or_null(host_lib, name);
201 XBT_ERROR("host %s already exits", name);
206 /* Note: ind_vm and vm_workstation point to the same elm object. */
207 msg_vm_t ind_vm = NULL;
208 void *ind_vm_workstation = NULL;
210 /* Ask the SIMIX layer to create the surf vm resource */
211 ind_vm_workstation = simcall_vm_create(name, ind_pm);
212 ind_vm = (msg_vm_t) __MSG_host_create(ind_vm_workstation);
214 XBT_DEBUG("A new VM (%s) has been created", name);
217 TRACE_msg_vm_create(name, ind_pm);
223 /** @brief Destroy a VM. Destroy the VM object from the simulation.
226 void MSG_vm_destroy(msg_vm_t vm)
228 /* First, terminate all processes on the VM if necessary */
229 if (MSG_vm_is_running(vm))
230 simcall_vm_shutdown(vm);
232 if (!MSG_vm_is_created(vm)) {
233 XBT_CRITICAL("shutdown the given VM before destroying it");
237 /* Then, destroy the VM object */
238 simcall_vm_destroy(vm);
240 __MSG_host_destroy(vm);
243 TRACE_msg_vm_end(vm);
248 /** @brief Start a vm (i.e., boot the guest operating system)
251 * If the VM cannot be started, an exception is generated.
254 void MSG_vm_start(msg_vm_t vm)
256 simcall_vm_start(vm);
259 TRACE_msg_vm_start(vm);
265 /** @brief Immediately kills all processes within the given VM. Any memory that they allocated will be leaked.
268 * FIXME: No extra delay occurs. If you want to simulate this too, you want to
269 * use a #MSG_process_sleep() or something. I'm not quite sure.
271 void MSG_vm_shutdown(msg_vm_t vm)
273 /* msg_vm_t equals to msg_host_t */
274 simcall_vm_shutdown(vm);
276 // #ifdef HAVE_TRACING
277 // TRACE_msg_vm_(vm);
283 /* We have two mailboxes. mbox is used to transfer migration data between
284 * source and destiantion PMs. mbox_ctl is used to detect the completion of a
285 * migration. The names of these mailboxes must not conflict with others. */
286 static inline char *get_mig_mbox_src_dst(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
288 return bprintf("__mbox_mig_src_dst:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
291 static inline char *get_mig_mbox_ctl(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
293 return bprintf("__mbox_mig_ctl:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
296 static inline char *get_mig_process_tx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
298 return bprintf("__pr_mig_tx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
301 static inline char *get_mig_process_rx_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name)
303 return bprintf("__pr_mig_rx:%s(%s-%s)", vm_name, src_pm_name, dst_pm_name);
306 static inline char *get_mig_task_name(const char *vm_name, const char *src_pm_name, const char *dst_pm_name, int stage)
308 return bprintf("__task_mig_stage%d:%s(%s-%s)", stage, vm_name, src_pm_name, dst_pm_name);
311 static int migration_rx_fun(int argc, char *argv[])
313 const char *pr_name = MSG_process_get_name(MSG_process_self());
314 const char *host_name = MSG_host_get_name(MSG_host_self());
316 XBT_DEBUG("mig: rx_start");
318 xbt_assert(argc == 4);
319 const char *vm_name = argv[1];
320 const char *src_pm_name = argv[2];
321 const char *dst_pm_name = argv[3];
322 msg_vm_t vm = MSG_get_host_by_name(vm_name);
323 msg_vm_t dst_pm = MSG_get_host_by_name(dst_pm_name);
327 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
328 char *mbox_ctl = get_mig_mbox_ctl(vm_name, src_pm_name, dst_pm_name);
329 char *finalize_task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 3);
332 msg_task_t task = NULL;
333 MSG_task_recv(&task, mbox);
335 if (strcmp(task->name, finalize_task_name) == 0)
338 MSG_task_destroy(task);
345 simcall_vm_migrate(vm, dst_pm);
346 simcall_vm_resume(vm);
349 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, 4);
351 msg_task_t task = MSG_task_create(task_name, 0, 0, NULL);
352 msg_error_t ret = MSG_task_send(task, mbox_ctl);
353 xbt_assert(ret == MSG_OK);
361 xbt_free(finalize_task_name);
363 XBT_DEBUG("mig: rx_done");
369 typedef struct dirty_page {
371 double prev_remaining;
373 } s_dirty_page, *dirty_page_t;
376 static void reset_dirty_pages(msg_vm_t vm)
378 msg_host_priv_t priv = msg_host_resource_priv(vm);
381 xbt_dict_cursor_t cursor = NULL;
382 dirty_page_t dp = NULL;
383 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
384 double remaining = MSG_task_get_remaining_computation(dp->task);
385 dp->prev_clock = MSG_get_clock();
386 dp->prev_remaining = remaining;
388 // XBT_INFO("%s@%s remaining %f", key, sg_host_name(vm), remaining);
392 static void start_dirty_page_tracking(msg_vm_t vm)
394 msg_host_priv_t priv = msg_host_resource_priv(vm);
395 priv->dp_enabled = 1;
397 reset_dirty_pages(vm);
400 static void stop_dirty_page_tracking(msg_vm_t vm)
402 msg_host_priv_t priv = msg_host_resource_priv(vm);
403 priv->dp_enabled = 0;
407 /* It might be natural that we define dp_rate for each task. But, we will also
408 * have to care about how each task behavior affects the memory update behavior
409 * at the operating system level. It may not be easy to model it with a simple algorithm. */
410 double calc_updated_pages(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
412 double computed = dp->prev_remaining - remaining;
413 double duration = clock - dp->prev_clock;
414 double updated = dp->task->dp_rate * computed;
416 XBT_INFO("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
417 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
418 XBT_INFO("%s@%s: updated %f bytes, %f Mbytes/s",
419 key, sg_host_name(vm), updated, updated / duration / 1000 / 1000);
425 double get_computed(char *key, msg_vm_t vm, dirty_page_t dp, double remaining, double clock)
427 double computed = dp->prev_remaining - remaining;
428 double duration = clock - dp->prev_clock;
430 XBT_DEBUG("%s@%s: computated %f ops (remaining %f -> %f) in %f secs (%f -> %f)",
431 key, sg_host_name(vm), computed, dp->prev_remaining, remaining, duration, dp->prev_clock, clock);
436 static double lookup_computed_flop_counts(msg_vm_t vm, int stage2_round_for_fancy_debug)
438 msg_host_priv_t priv = msg_host_resource_priv(vm);
442 xbt_dict_cursor_t cursor = NULL;
443 dirty_page_t dp = NULL;
444 xbt_dict_foreach(priv->dp_objs, cursor, key, dp) {
445 double remaining = MSG_task_get_remaining_computation(dp->task);
446 double clock = MSG_get_clock();
448 // total += calc_updated_pages(key, vm, dp, remaining, clock);
449 total += get_computed(key, vm, dp, remaining, clock);
451 dp->prev_remaining = remaining;
452 dp->prev_clock = clock;
455 total += priv->dp_updated_by_deleted_tasks;
457 XBT_INFO("mig-stage2.%d: computed %f flop_counts (including %f by deleted tasks)",
458 stage2_round_for_fancy_debug,
459 total, priv->dp_updated_by_deleted_tasks);
463 priv->dp_updated_by_deleted_tasks = 0;
469 // TODO Is this code redundant with the information provided by
470 // msg_process_t MSG_process_create(const char *name, xbt_main_func_t code, void *data, msg_host_t host)
471 void MSG_host_add_task(msg_host_t host, msg_task_t task)
473 msg_host_priv_t priv = msg_host_resource_priv(host);
474 double remaining = MSG_task_get_remaining_computation(task);
475 char *key = bprintf("%s-%lld", task->name, task->counter);
477 dirty_page_t dp = xbt_new0(s_dirty_page, 1);
480 /* It should be okay that we add a task onto a migrating VM. */
481 if (priv->dp_enabled) {
482 dp->prev_clock = MSG_get_clock();
483 dp->prev_remaining = remaining;
486 xbt_assert(xbt_dict_get_or_null(priv->dp_objs, key) == NULL);
487 xbt_dict_set(priv->dp_objs, key, dp, NULL);
488 XBT_DEBUG("add %s on %s (remaining %f, dp_enabled %d)", key, sg_host_name(host), remaining, priv->dp_enabled);
493 void MSG_host_del_task(msg_host_t host, msg_task_t task)
495 msg_host_priv_t priv = msg_host_resource_priv(host);
497 char *key = bprintf("%s-%lld", task->name, task->counter);
499 dirty_page_t dp = xbt_dict_get_or_null(priv->dp_objs, key);
500 xbt_assert(dp->task == task);
502 /* If we are in the middle of dirty page tracking, we record how much
503 * computaion has been done until now, and keep the information for the
504 * lookup_() function that will called soon. */
505 if (priv->dp_enabled) {
506 double remaining = MSG_task_get_remaining_computation(task);
507 double clock = MSG_get_clock();
508 // double updated = calc_updated_pages(key, host, dp, remaining, clock);
509 double updated = get_computed(key, host, dp, remaining, clock);
511 priv->dp_updated_by_deleted_tasks += updated;
514 xbt_dict_remove(priv->dp_objs, key);
517 XBT_DEBUG("del %s on %s", key, sg_host_name(host));
523 static void send_migration_data(const char *vm_name, const char *src_pm_name, const char *dst_pm_name,
524 double size, char *mbox, int stage, int stage2_round, double mig_speed)
526 char *task_name = get_mig_task_name(vm_name, src_pm_name, dst_pm_name, stage);
527 msg_task_t task = MSG_task_create(task_name, 0, size, NULL);
531 ret = MSG_task_send_bounded(task, mbox, mig_speed);
533 ret = MSG_task_send(task, mbox);
534 xbt_assert(ret == MSG_OK);
537 XBT_INFO("mig-stage%d.%d: sent %f", stage, stage2_round, size);
539 XBT_INFO("mig-stage%d: sent %f", stage, size);
545 static int migration_tx_fun(int argc, char *argv[])
547 const char *pr_name = MSG_process_get_name(MSG_process_self());
548 const char *host_name = MSG_host_get_name(MSG_host_self());
550 XBT_DEBUG("mig: tx_start");
552 xbt_assert(argc == 4);
553 const char *vm_name = argv[1];
554 const char *src_pm_name = argv[2];
555 const char *dst_pm_name = argv[3];
556 msg_vm_t vm = MSG_get_host_by_name(vm_name);
559 s_ws_params_t params;
560 simcall_host_get_params(vm, ¶ms);
561 const long ramsize = params.ramsize;
562 const long devsize = params.devsize;
563 const int skip_stage2 = params.skip_stage2;
564 const double max_downtime = params.max_downtime;
565 const double dp_rate = params.dp_rate;
566 const double dp_cap = params.dp_cap;
567 const double mig_speed = params.mig_speed;
568 double remaining_size = ramsize + devsize;
569 double threshold = max_downtime * 125 * 1000 * 1000;
573 XBT_WARN("migrate a VM, but ramsize is zero");
575 char *mbox = get_mig_mbox_src_dst(vm_name, src_pm_name, dst_pm_name);
577 XBT_INFO("mig-stage1: remaining_size %f", remaining_size);
579 /* Stage1: send all memory pages to the destination. */
580 start_dirty_page_tracking(vm);
582 send_migration_data(vm_name, src_pm_name, dst_pm_name, ramsize, mbox, 1, 0, mig_speed);
584 remaining_size -= ramsize;
588 /* Stage2: send update pages iteratively until the size of remaining states
589 * becomes smaller than the threshold value. */
592 if (max_downtime == 0) {
593 XBT_WARN("no max_downtime parameter, skip stage2");
598 int stage2_round = 0;
600 // long updated_size = lookup_dirty_pages(vm);
601 double updated_size = lookup_computed_flop_counts(vm, stage2_round) * dp_rate;
602 if (updated_size > dp_cap) {
603 XBT_INFO("mig-stage2.%d: %f bytes updated, but cap it with the working set size %f",
604 stage2_round, updated_size, dp_cap);
605 updated_size = dp_cap;
608 remaining_size += updated_size;
610 XBT_INFO("mig-stage2.%d: remaining_size %f (%s threshold %f)", stage2_round,
611 remaining_size, (remaining_size < threshold) ? "<" : ">", threshold);
613 if (remaining_size < threshold)
616 send_migration_data(vm_name, src_pm_name, dst_pm_name, updated_size, mbox, 2, stage2_round, mig_speed);
618 remaining_size -= updated_size;
624 /* Stage3: stop the VM and copy the rest of states. */
625 XBT_INFO("mig-stage3: remaining_size %f", remaining_size);
626 simcall_vm_suspend(vm);
627 stop_dirty_page_tracking(vm);
629 send_migration_data(vm_name, src_pm_name, dst_pm_name, remaining_size, mbox, 3, 0, mig_speed);
633 XBT_DEBUG("mig: tx_done");
640 static void do_migration(msg_vm_t vm, msg_host_t src_pm, msg_host_t dst_pm)
642 char *mbox_ctl = get_mig_mbox_ctl(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
645 char *pr_name = get_mig_process_rx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
647 char **argv = xbt_new(char *, nargvs);
648 argv[0] = xbt_strdup(pr_name);
649 argv[1] = xbt_strdup(sg_host_name(vm));
650 argv[2] = xbt_strdup(sg_host_name(src_pm));
651 argv[3] = xbt_strdup(sg_host_name(dst_pm));
654 msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_rx_fun, NULL, dst_pm, nargvs - 1, argv);
660 char *pr_name = get_mig_process_tx_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm));
662 char **argv = xbt_new(char *, nargvs);
663 argv[0] = xbt_strdup(pr_name);
664 argv[1] = xbt_strdup(sg_host_name(vm));
665 argv[2] = xbt_strdup(sg_host_name(src_pm));
666 argv[3] = xbt_strdup(sg_host_name(dst_pm));
669 msg_process_t pr = MSG_process_create_with_arguments(pr_name, migration_tx_fun, NULL, src_pm, nargvs - 1, argv);
674 /* wait until the migration have finished */
676 msg_task_t task = NULL;
677 msg_error_t ret = MSG_task_recv(&task, mbox_ctl);
678 xbt_assert(ret == MSG_OK);
680 char *expected_task_name = get_mig_task_name(sg_host_name(vm), sg_host_name(src_pm), sg_host_name(dst_pm), 4);
681 xbt_assert(strcmp(task->name, expected_task_name) == 0);
682 xbt_free(expected_task_name);
689 /** @brief Migrate the VM to the given host.
692 * FIXME: No migration cost occurs. If you want to simulate this too, you want to use a
693 * MSG_task_send() before or after, depending on whether you want to do cold or hot
696 void MSG_vm_migrate(msg_vm_t vm, msg_host_t new_pm)
699 * - One approach is ...
700 * We first create a new VM (i.e., destination VM) on the destination
701 * physical host. The destination VM will receive the state of the source
702 * VM over network. We will finally destroy the source VM.
703 * - This behavior is similar to the way of migration in the real world.
704 * Even before a migration is completed, we will see a destination VM,
705 * consuming resources.
706 * - We have to relocate all processes. The existing process migraion code
707 * will work for this?
708 * - The name of the VM is a somewhat unique ID in the code. It is tricky
709 * for the destination VM?
711 * - Another one is ...
712 * We update the information of the given VM to place it to the destination
715 * The second one would be easier.
719 msg_host_t old_pm = simcall_vm_get_pm(vm);
721 if (simcall_vm_get_state(vm) != SURF_VM_STATE_RUNNING)
722 THROWF(vm_error, 0, "VM(%s) is not running", sg_host_name(vm));
724 do_migration(vm, old_pm, new_pm);
728 XBT_DEBUG("VM(%s) moved from PM(%s) to PM(%s)", vm->key, old_pm->key, new_pm->key);
731 TRACE_msg_vm_change_host(vm, old_pm, new_pm);
736 /** @brief Immediately suspend the execution of all processes within the given VM.
739 * This function stops the exection of the VM. All the processes on this VM
740 * will pause. The state of the VM is perserved. We can later resume it again.
742 * No suspension cost occurs.
744 void MSG_vm_suspend(msg_vm_t vm)
746 simcall_vm_suspend(vm);
748 XBT_DEBUG("vm_suspend done");
751 TRACE_msg_vm_suspend(vm);
756 /** @brief Resume the execution of the VM. All processes on the VM run again.
759 * No resume cost occurs.
761 void MSG_vm_resume(msg_vm_t vm)
763 simcall_vm_resume(vm);
766 TRACE_msg_vm_resume(vm);
771 /** @brief Immediately save the execution of all processes within the given VM.
774 * This function stops the exection of the VM. All the processes on this VM
775 * will pause. The state of the VM is perserved. We can later resume it again.
777 * FIXME: No suspension cost occurs. If you want to simulate this too, you want to
778 * use a \ref MSG_file_write() before or after, depending on the exact semantic
781 void MSG_vm_save(msg_vm_t vm)
785 TRACE_msg_vm_save(vm);
789 /** @brief Restore the execution of the VM. All processes on the VM run again.
792 * FIXME: No restore cost occurs. If you want to simulate this too, you want to
793 * use a \ref MSG_file_read() before or after, depending on the exact semantic
794 * of VM restore to you.
796 void MSG_vm_restore(msg_vm_t vm)
798 simcall_vm_restore(vm);
801 TRACE_msg_vm_restore(vm);
808 /** @brief Get the physical host of a given VM.
811 msg_host_t MSG_vm_get_pm(msg_vm_t vm)
813 return simcall_vm_get_pm(vm);